comp.lang.idl-pvwave archive
Messages from Usenet group comp.lang.idl-pvwave, compiled by Paulo Penteado
Home » Public Forums » archive » Re: Speed does matter
Show: Today's Messages :: Show Polls :: Message Navigator
E-mail to friend  		Return to the default flat view Create a new topic Submit Reply
Re: Speed does matter [message #91782 is a reply to message #91780] Wed, 26 August 2015 01:41 Go to previous messageGo to previous message
Kallisthène is currently offline  Kallisthène
Messages: 15
Registered: October 2012
Junior Member
Le mardi 25 août 2015 23:35:11 UTC+2, Michael Galloy a écrit :
> On 8/25/15 7:03 AM, Kallisthène wrote:
>> Le mardi 25 août 2015 10:32:22 UTC+2, superchromix a écrit :
>>> Nice comparison. On my desktop, running IDL 8.5 and Matlab 2010b,
>>> I get the following results:
>>>
>>> IDL: % Time elapsed: 6.4400001 seconds.
>>>
>>> MATLAB 2010b: Elapsed time is 0.594112 seconds.
>>>
>>> Again, MATLAB for loops run faster by a factor of ~10.
>>
>> Great idea,
>>
>>
>> I ran also these codes, with bigger numbers, 1000 instead of 500 for
>> each dimension and with Matlab 2015a I got 6.3 secondes to compare
>> with 102 secondes on my IDL 8.5. My desktop runs 4 cores and has 12
>> Go RAM and it runs faster with Matlab by a factor of 16 ! Damn.
>>
>
> For what it's worth, I got about a 5x speedup of IDL over Python. I used
> 1000 elements for each dimension:
>
> $ python mg_loopspeed.py
> 222.15 secs
>
> Code is:
>
> import numpy as np
> import time
>
> nx = 1000
> ny = 1000
> nz = 1000
>
> arr = np.zeros((nx, ny, nz))
>
> tic = time.clock()
>
> for z in xrange(nz):
> for y in xrange(ny):
> for x in xrange(nx):
> arr[x, y, z] = 1.0
>
> toc = time.clock()
> print '%0.2f secs' % (toc - tic)
>
> For comparison, my IDL time:
>
> $ idl -e ".run mg_loopspeed"
> IDL Version 8.5, Mac OS X (darwin x86_64 m64).
> (c) 2015, Exelis Visual Information Solutions, Inc., a subsidiary of
> Harris Corporation.
>
> % Compiled module: $MAIN$.
> % Compiled module: TIC.
> Elapsed Time: 47.486635
>
> Mike
> --
> Michael Galloy
> www.michaelgalloy.com
> Modern IDL: A Guide to IDL Programming (http://modernidl.idldev.com)


That's right, I got similar results on my Winpython 3.4. The reason seems to lie in the Just In Time compilation heavily used by Matlab (and Julia as well). To check it I used the "feature JIT off" undocumented Matlab command to see how it would fare.
And I got 48 secondes instead of the previous 6.3 secondes, an impressive change which puts Matlab un-JITed in the same league as IDL (102 secondes).
It is likely that's the reason.
To go further on this point with Python we'd need to check the efficiency of the Pypy Python interpreter and JIT compiler.

Seems to me that the work necessary to develop a JIT compiler for IDL is much heavier than compiling IDL against MKL.


Best

Best

Report message to a moderator

[Message index]
 
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Read Message
Previous Topic: Any way to initialize mapping in IDL with WKT or a Proj.4 string?
Next Topic: getting uniform cadence data

-=] Back to Top [=-
[ Syndicate this forum (XML) ] [ RSS ] [ PDF ]

Current Time: Wed Oct 08 19:31:18 PDT 2025

Total time taken to generate the page: 0.00533 seconds
.:: Contact :: Home ::.

Powered by: FUDforum 3.0.7.
Copyright ©2001-2013 FUDforum Bulletin Board Software